1 //===- FunctionImport.cpp - ThinLTO Summary-based Function Import ---------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements Function import based on summaries.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "llvm/Transforms/IPO/FunctionImport.h"
15 
16 #include "llvm/ADT/SmallVector.h"
17 #include "llvm/ADT/Statistic.h"
18 #include "llvm/ADT/StringSet.h"
19 #include "llvm/ADT/Triple.h"
20 #include "llvm/IR/AutoUpgrade.h"
21 #include "llvm/IR/DiagnosticPrinter.h"
22 #include "llvm/IR/IntrinsicInst.h"
23 #include "llvm/IR/Module.h"
24 #include "llvm/IRReader/IRReader.h"
25 #include "llvm/Linker/Linker.h"
26 #include "llvm/Object/IRObjectFile.h"
27 #include "llvm/Object/ModuleSummaryIndexObjectFile.h"
28 #include "llvm/Support/CommandLine.h"
29 #include "llvm/Support/Debug.h"
30 #include "llvm/Support/SourceMgr.h"
31 #include "llvm/Transforms/IPO/Internalize.h"
32 #include "llvm/Transforms/Utils/FunctionImportUtils.h"
33 
34 #define DEBUG_TYPE "function-import"
35 
36 using namespace llvm;
37 
38 STATISTIC(NumImported, "Number of functions imported");
39 
40 /// Limit on instruction count of imported functions.
41 static cl::opt<unsigned> ImportInstrLimit(
42     "import-instr-limit", cl::init(100), cl::Hidden, cl::value_desc("N"),
43     cl::desc("Only import functions with less than N instructions"));
44 
45 static cl::opt<float>
46     ImportInstrFactor("import-instr-evolution-factor", cl::init(0.7),
47                       cl::Hidden, cl::value_desc("x"),
48                       cl::desc("As we import functions, multiply the "
49                                "`import-instr-limit` threshold by this factor "
50                                "before processing newly imported functions"));
51 
52 static cl::opt<bool> PrintImports("print-imports", cl::init(false), cl::Hidden,
53                                   cl::desc("Print imported functions"));
54 
55 // Temporary allows the function import pass to disable always linking
56 // referenced discardable symbols.
57 static cl::opt<bool>
58     DontForceImportReferencedDiscardableSymbols("disable-force-link-odr",
59                                                 cl::init(false), cl::Hidden);
60 
61 static cl::opt<bool> EnableImportMetadata(
62     "enable-import-metadata", cl::init(
63 #if !defined(NDEBUG)
64                                   true /*Enabled with asserts.*/
65 #else
66                                   false
67 #endif
68                                   ),
69     cl::Hidden, cl::desc("Enable import metadata like 'thinlto_src_module'"));
70 
71 // Load lazily a module from \p FileName in \p Context.
72 static std::unique_ptr<Module> loadFile(const std::string &FileName,
73                                         LLVMContext &Context) {
74   SMDiagnostic Err;
75   DEBUG(dbgs() << "Loading '" << FileName << "'\n");
76   // Metadata isn't loaded until functions are imported, to minimize
77   // the memory overhead.
78   std::unique_ptr<Module> Result =
79       getLazyIRFileModule(FileName, Err, Context,
80                           /* ShouldLazyLoadMetadata = */ true);
81   if (!Result) {
82     Err.print("function-import", errs());
83     report_fatal_error("Abort");
84   }
85 
86   return Result;
87 }
88 
89 namespace {
90 
91 // Return true if the Summary describes a GlobalValue that can be externally
92 // referenced, i.e. it does not need renaming (linkage is not local) or renaming
93 // is possible (does not have a section for instance).
94 static bool canBeExternallyReferenced(const GlobalValueSummary &Summary) {
95   if (!Summary.needsRenaming())
96     return true;
97 
98   if (Summary.hasSection())
99     // Can't rename a global that needs renaming if has a section.
100     return false;
101 
102   return true;
103 }
104 
105 // Return true if \p GUID describes a GlobalValue that can be externally
106 // referenced, i.e. it does not need renaming (linkage is not local) or
107 // renaming is possible (does not have a section for instance).
108 static bool canBeExternallyReferenced(const ModuleSummaryIndex &Index,
109                                       GlobalValue::GUID GUID) {
110   auto Summaries = Index.findGlobalValueSummaryList(GUID);
111   if (Summaries == Index.end())
112     return true;
113   if (Summaries->second.size() != 1)
114     // If there are multiple globals with this GUID, then we know it is
115     // not a local symbol, and it is necessarily externally referenced.
116     return true;
117 
118   // We don't need to check for the module path, because if it can't be
119   // externally referenced and we call it, it is necessarilly in the same
120   // module
121   return canBeExternallyReferenced(**Summaries->second.begin());
122 }
123 
124 // Return true if the global described by \p Summary can be imported in another
125 // module.
126 static bool eligibleForImport(const ModuleSummaryIndex &Index,
127                               const GlobalValueSummary &Summary) {
128   if (!canBeExternallyReferenced(Summary))
129     // Can't import a global that needs renaming if has a section for instance.
130     // FIXME: we may be able to import it by copying it without promotion.
131     return false;
132 
133   // Don't import functions that are not viable to inline.
134   if (Summary.isNotViableToInline())
135     return false;
136 
137   // Check references (and potential calls) in the same module. If the current
138   // value references a global that can't be externally referenced it is not
139   // eligible for import.
140   bool AllRefsCanBeExternallyReferenced =
141       llvm::all_of(Summary.refs(), [&](const ValueInfo &VI) {
142         return canBeExternallyReferenced(Index, VI.getGUID());
143       });
144   if (!AllRefsCanBeExternallyReferenced)
145     return false;
146 
147   if (auto *FuncSummary = dyn_cast<FunctionSummary>(&Summary)) {
148     bool AllCallsCanBeExternallyReferenced = llvm::all_of(
149         FuncSummary->calls(), [&](const FunctionSummary::EdgeTy &Edge) {
150           return canBeExternallyReferenced(Index, Edge.first.getGUID());
151         });
152     if (!AllCallsCanBeExternallyReferenced)
153       return false;
154   }
155   return true;
156 }
157 
158 /// Given a list of possible callee implementation for a call site, select one
159 /// that fits the \p Threshold.
160 ///
161 /// FIXME: select "best" instead of first that fits. But what is "best"?
162 /// - The smallest: more likely to be inlined.
163 /// - The one with the least outgoing edges (already well optimized).
164 /// - One from a module already being imported from in order to reduce the
165 ///   number of source modules parsed/linked.
166 /// - One that has PGO data attached.
167 /// - [insert you fancy metric here]
168 static const GlobalValueSummary *
169 selectCallee(const ModuleSummaryIndex &Index,
170              const GlobalValueSummaryList &CalleeSummaryList,
171              unsigned Threshold) {
172   auto It = llvm::find_if(
173       CalleeSummaryList,
174       [&](const std::unique_ptr<GlobalValueSummary> &SummaryPtr) {
175         auto *GVSummary = SummaryPtr.get();
176         if (GlobalValue::isInterposableLinkage(GVSummary->linkage()))
177           // There is no point in importing these, we can't inline them
178           return false;
179         if (auto *AS = dyn_cast<AliasSummary>(GVSummary)) {
180           GVSummary = &AS->getAliasee();
181           // Alias can't point to "available_externally". However when we import
182           // linkOnceODR the linkage does not change. So we import the alias
183           // and aliasee only in this case.
184           // FIXME: we should import alias as available_externally *function*,
185           // the destination module does need to know it is an alias.
186           if (!GlobalValue::isLinkOnceODRLinkage(GVSummary->linkage()))
187             return false;
188         }
189 
190         auto *Summary = cast<FunctionSummary>(GVSummary);
191 
192         if (Summary->instCount() > Threshold)
193           return false;
194 
195         if (!eligibleForImport(Index, *Summary))
196           return false;
197 
198         return true;
199       });
200   if (It == CalleeSummaryList.end())
201     return nullptr;
202 
203   return cast<GlobalValueSummary>(It->get());
204 }
205 
206 /// Return the summary for the function \p GUID that fits the \p Threshold, or
207 /// null if there's no match.
208 static const GlobalValueSummary *selectCallee(GlobalValue::GUID GUID,
209                                               unsigned Threshold,
210                                               const ModuleSummaryIndex &Index) {
211   auto CalleeSummaryList = Index.findGlobalValueSummaryList(GUID);
212   if (CalleeSummaryList == Index.end())
213     return nullptr; // This function does not have a summary
214   return selectCallee(Index, CalleeSummaryList->second, Threshold);
215 }
216 
217 /// Mark the global \p GUID as export by module \p ExportModulePath if found in
218 /// this module. If it is a GlobalVariable, we also mark any referenced global
219 /// in the current module as exported.
220 static void exportGlobalInModule(const ModuleSummaryIndex &Index,
221                                  StringRef ExportModulePath,
222                                  GlobalValue::GUID GUID,
223                                  FunctionImporter::ExportSetTy &ExportList) {
224   auto FindGlobalSummaryInModule =
225       [&](GlobalValue::GUID GUID) -> GlobalValueSummary *{
226         auto SummaryList = Index.findGlobalValueSummaryList(GUID);
227         if (SummaryList == Index.end())
228           // This global does not have a summary, it is not part of the ThinLTO
229           // process
230           return nullptr;
231         auto SummaryIter = llvm::find_if(
232             SummaryList->second,
233             [&](const std::unique_ptr<GlobalValueSummary> &Summary) {
234               return Summary->modulePath() == ExportModulePath;
235             });
236         if (SummaryIter == SummaryList->second.end())
237           return nullptr;
238         return SummaryIter->get();
239       };
240 
241   auto *Summary = FindGlobalSummaryInModule(GUID);
242   if (!Summary)
243     return;
244   // We found it in the current module, mark as exported
245   ExportList.insert(GUID);
246 
247   auto GVS = dyn_cast<GlobalVarSummary>(Summary);
248   if (!GVS)
249     return;
250   // FunctionImportGlobalProcessing::doPromoteLocalToGlobal() will always
251   // trigger importing  the initializer for `constant unnamed addr` globals that
252   // are referenced. We conservatively export all the referenced symbols for
253   // every global to workaround this, so that the ExportList is accurate.
254   // FIXME: with a "isConstant" flag in the summary we could be more targetted.
255   for (auto &Ref : GVS->refs()) {
256     auto GUID = Ref.getGUID();
257     auto *RefSummary = FindGlobalSummaryInModule(GUID);
258     if (RefSummary)
259       // Found a ref in the current module, mark it as exported
260       ExportList.insert(GUID);
261   }
262 }
263 
264 using EdgeInfo = std::pair<const FunctionSummary *, unsigned /* Threshold */>;
265 
266 /// Compute the list of functions to import for a given caller. Mark these
267 /// imported functions and the symbols they reference in their source module as
268 /// exported from their source module.
269 static void computeImportForFunction(
270     const FunctionSummary &Summary, const ModuleSummaryIndex &Index,
271     unsigned Threshold, const GVSummaryMapTy &DefinedGVSummaries,
272     SmallVectorImpl<EdgeInfo> &Worklist,
273     FunctionImporter::ImportMapTy &ImportList,
274     StringMap<FunctionImporter::ExportSetTy> *ExportLists = nullptr) {
275   for (auto &Edge : Summary.calls()) {
276     auto GUID = Edge.first.getGUID();
277     DEBUG(dbgs() << " edge -> " << GUID << " Threshold:" << Threshold << "\n");
278 
279     if (DefinedGVSummaries.count(GUID)) {
280       DEBUG(dbgs() << "ignored! Target already in destination module.\n");
281       continue;
282     }
283 
284     auto *CalleeSummary = selectCallee(GUID, Threshold, Index);
285     if (!CalleeSummary) {
286       DEBUG(dbgs() << "ignored! No qualifying callee with summary found.\n");
287       continue;
288     }
289     // "Resolve" the summary, traversing alias,
290     const FunctionSummary *ResolvedCalleeSummary;
291     if (isa<AliasSummary>(CalleeSummary)) {
292       ResolvedCalleeSummary = cast<FunctionSummary>(
293           &cast<AliasSummary>(CalleeSummary)->getAliasee());
294       assert(
295           GlobalValue::isLinkOnceODRLinkage(ResolvedCalleeSummary->linkage()) &&
296           "Unexpected alias to a non-linkonceODR in import list");
297     } else
298       ResolvedCalleeSummary = cast<FunctionSummary>(CalleeSummary);
299 
300     assert(ResolvedCalleeSummary->instCount() <= Threshold &&
301            "selectCallee() didn't honor the threshold");
302 
303     auto ExportModulePath = ResolvedCalleeSummary->modulePath();
304     auto &ProcessedThreshold = ImportList[ExportModulePath][GUID];
305     /// Since the traversal of the call graph is DFS, we can revisit a function
306     /// a second time with a higher threshold. In this case, it is added back to
307     /// the worklist with the new threshold.
308     if (ProcessedThreshold && ProcessedThreshold >= Threshold) {
309       DEBUG(dbgs() << "ignored! Target was already seen with Threshold "
310                    << ProcessedThreshold << "\n");
311       continue;
312     }
313     // Mark this function as imported in this module, with the current Threshold
314     ProcessedThreshold = Threshold;
315 
316     // Make exports in the source module.
317     if (ExportLists) {
318       auto &ExportList = (*ExportLists)[ExportModulePath];
319       ExportList.insert(GUID);
320       // Mark all functions and globals referenced by this function as exported
321       // to the outside if they are defined in the same source module.
322       for (auto &Edge : ResolvedCalleeSummary->calls()) {
323         auto CalleeGUID = Edge.first.getGUID();
324         exportGlobalInModule(Index, ExportModulePath, CalleeGUID, ExportList);
325       }
326       for (auto &Ref : ResolvedCalleeSummary->refs()) {
327         auto GUID = Ref.getGUID();
328         exportGlobalInModule(Index, ExportModulePath, GUID, ExportList);
329       }
330     }
331 
332     // Insert the newly imported function to the worklist.
333     Worklist.push_back(std::make_pair(ResolvedCalleeSummary, Threshold));
334   }
335 }
336 
337 /// Given the list of globals defined in a module, compute the list of imports
338 /// as well as the list of "exports", i.e. the list of symbols referenced from
339 /// another module (that may require promotion).
340 static void ComputeImportForModule(
341     const GVSummaryMapTy &DefinedGVSummaries, const ModuleSummaryIndex &Index,
342     FunctionImporter::ImportMapTy &ImportList,
343     StringMap<FunctionImporter::ExportSetTy> *ExportLists = nullptr) {
344   // Worklist contains the list of function imported in this module, for which
345   // we will analyse the callees and may import further down the callgraph.
346   SmallVector<EdgeInfo, 128> Worklist;
347 
348   // Populate the worklist with the import for the functions in the current
349   // module
350   for (auto &GVSummary : DefinedGVSummaries) {
351     auto *Summary = GVSummary.second;
352     if (auto *AS = dyn_cast<AliasSummary>(Summary))
353       Summary = &AS->getAliasee();
354     auto *FuncSummary = dyn_cast<FunctionSummary>(Summary);
355     if (!FuncSummary)
356       // Skip import for global variables
357       continue;
358     DEBUG(dbgs() << "Initalize import for " << GVSummary.first << "\n");
359     computeImportForFunction(*FuncSummary, Index, ImportInstrLimit,
360                              DefinedGVSummaries, Worklist, ImportList,
361                              ExportLists);
362   }
363 
364   while (!Worklist.empty()) {
365     auto FuncInfo = Worklist.pop_back_val();
366     auto *Summary = FuncInfo.first;
367     auto Threshold = FuncInfo.second;
368 
369     // Process the newly imported functions and add callees to the worklist.
370     // Adjust the threshold
371     Threshold = Threshold * ImportInstrFactor;
372 
373     computeImportForFunction(*Summary, Index, Threshold, DefinedGVSummaries,
374                              Worklist, ImportList, ExportLists);
375   }
376 }
377 
378 } // anonymous namespace
379 
380 /// Compute all the import and export for every module using the Index.
381 void llvm::ComputeCrossModuleImport(
382     const ModuleSummaryIndex &Index,
383     const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries,
384     StringMap<FunctionImporter::ImportMapTy> &ImportLists,
385     StringMap<FunctionImporter::ExportSetTy> &ExportLists) {
386   // For each module that has function defined, compute the import/export lists.
387   for (auto &DefinedGVSummaries : ModuleToDefinedGVSummaries) {
388     auto &ImportList = ImportLists[DefinedGVSummaries.first()];
389     DEBUG(dbgs() << "Computing import for Module '"
390                  << DefinedGVSummaries.first() << "'\n");
391     ComputeImportForModule(DefinedGVSummaries.second, Index, ImportList,
392                            &ExportLists);
393   }
394 
395 #ifndef NDEBUG
396   DEBUG(dbgs() << "Import/Export lists for " << ImportLists.size()
397                << " modules:\n");
398   for (auto &ModuleImports : ImportLists) {
399     auto ModName = ModuleImports.first();
400     auto &Exports = ExportLists[ModName];
401     DEBUG(dbgs() << "* Module " << ModName << " exports " << Exports.size()
402                  << " functions. Imports from " << ModuleImports.second.size()
403                  << " modules.\n");
404     for (auto &Src : ModuleImports.second) {
405       auto SrcModName = Src.first();
406       DEBUG(dbgs() << " - " << Src.second.size() << " functions imported from "
407                    << SrcModName << "\n");
408     }
409   }
410 #endif
411 }
412 
413 /// Compute all the imports for the given module in the Index.
414 void llvm::ComputeCrossModuleImportForModule(
415     StringRef ModulePath, const ModuleSummaryIndex &Index,
416     FunctionImporter::ImportMapTy &ImportList) {
417 
418   // Collect the list of functions this module defines.
419   // GUID -> Summary
420   GVSummaryMapTy FunctionSummaryMap;
421   Index.collectDefinedFunctionsForModule(ModulePath, FunctionSummaryMap);
422 
423   // Compute the import list for this module.
424   DEBUG(dbgs() << "Computing import for Module '" << ModulePath << "'\n");
425   ComputeImportForModule(FunctionSummaryMap, Index, ImportList);
426 
427 #ifndef NDEBUG
428   DEBUG(dbgs() << "* Module " << ModulePath << " imports from "
429                << ImportList.size() << " modules.\n");
430   for (auto &Src : ImportList) {
431     auto SrcModName = Src.first();
432     DEBUG(dbgs() << " - " << Src.second.size() << " functions imported from "
433                  << SrcModName << "\n");
434   }
435 #endif
436 }
437 
438 /// Compute the set of summaries needed for a ThinLTO backend compilation of
439 /// \p ModulePath.
440 void llvm::gatherImportedSummariesForModule(
441     StringRef ModulePath,
442     const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries,
443     const FunctionImporter::ImportMapTy &ImportList,
444     std::map<std::string, GVSummaryMapTy> &ModuleToSummariesForIndex) {
445   // Include all summaries from the importing module.
446   ModuleToSummariesForIndex[ModulePath] =
447       ModuleToDefinedGVSummaries.lookup(ModulePath);
448   // Include summaries for imports.
449   for (auto &ILI : ImportList) {
450     auto &SummariesForIndex = ModuleToSummariesForIndex[ILI.first()];
451     const auto &DefinedGVSummaries =
452         ModuleToDefinedGVSummaries.lookup(ILI.first());
453     for (auto &GI : ILI.second) {
454       const auto &DS = DefinedGVSummaries.find(GI.first);
455       assert(DS != DefinedGVSummaries.end() &&
456              "Expected a defined summary for imported global value");
457       SummariesForIndex[GI.first] = DS->second;
458     }
459   }
460 }
461 
462 /// Emit the files \p ModulePath will import from into \p OutputFilename.
463 std::error_code
464 llvm::EmitImportsFiles(StringRef ModulePath, StringRef OutputFilename,
465                        const FunctionImporter::ImportMapTy &ModuleImports) {
466   std::error_code EC;
467   raw_fd_ostream ImportsOS(OutputFilename, EC, sys::fs::OpenFlags::F_None);
468   if (EC)
469     return EC;
470   for (auto &ILI : ModuleImports)
471     ImportsOS << ILI.first() << "\n";
472   return std::error_code();
473 }
474 
475 /// Fixup WeakForLinker linkages in \p TheModule based on summary analysis.
476 void llvm::thinLTOResolveWeakForLinkerModule(
477     Module &TheModule, const GVSummaryMapTy &DefinedGlobals) {
478   auto updateLinkage = [&](GlobalValue &GV) {
479     if (!GlobalValue::isWeakForLinker(GV.getLinkage()))
480       return;
481     // See if the global summary analysis computed a new resolved linkage.
482     const auto &GS = DefinedGlobals.find(GV.getGUID());
483     if (GS == DefinedGlobals.end())
484       return;
485     auto NewLinkage = GS->second->linkage();
486     if (NewLinkage == GV.getLinkage())
487       return;
488     DEBUG(dbgs() << "ODR fixing up linkage for `" << GV.getName() << "` from "
489                  << GV.getLinkage() << " to " << NewLinkage << "\n");
490     GV.setLinkage(NewLinkage);
491     // Remove functions converted to available_externally from comdats,
492     // as this is a declaration for the linker, and will be dropped eventually.
493     // It is illegal for comdats to contain declarations.
494     auto *GO = dyn_cast_or_null<GlobalObject>(&GV);
495     if (GO && GO->isDeclarationForLinker() && GO->hasComdat()) {
496       assert(GO->hasAvailableExternallyLinkage() &&
497              "Expected comdat on definition (possibly available external)");
498       GO->setComdat(nullptr);
499     }
500   };
501 
502   // Process functions and global now
503   for (auto &GV : TheModule)
504     updateLinkage(GV);
505   for (auto &GV : TheModule.globals())
506     updateLinkage(GV);
507   for (auto &GV : TheModule.aliases())
508     updateLinkage(GV);
509 }
510 
511 /// Run internalization on \p TheModule based on symmary analysis.
512 void llvm::thinLTOInternalizeModule(Module &TheModule,
513                                     const GVSummaryMapTy &DefinedGlobals) {
514   // Parse inline ASM and collect the list of symbols that are not defined in
515   // the current module.
516   StringSet<> AsmUndefinedRefs;
517   object::IRObjectFile::CollectAsmUndefinedRefs(
518       Triple(TheModule.getTargetTriple()), TheModule.getModuleInlineAsm(),
519       [&AsmUndefinedRefs](StringRef Name, object::BasicSymbolRef::Flags Flags) {
520         if (Flags & object::BasicSymbolRef::SF_Undefined)
521           AsmUndefinedRefs.insert(Name);
522       });
523 
524   // Declare a callback for the internalize pass that will ask for every
525   // candidate GlobalValue if it can be internalized or not.
526   auto MustPreserveGV = [&](const GlobalValue &GV) -> bool {
527     // Can't be internalized if referenced in inline asm.
528     if (AsmUndefinedRefs.count(GV.getName()))
529       return true;
530 
531     // Lookup the linkage recorded in the summaries during global analysis.
532     const auto &GS = DefinedGlobals.find(GV.getGUID());
533     GlobalValue::LinkageTypes Linkage;
534     if (GS == DefinedGlobals.end()) {
535       // Must have been promoted (possibly conservatively). Find original
536       // name so that we can access the correct summary and see if it can
537       // be internalized again.
538       // FIXME: Eventually we should control promotion instead of promoting
539       // and internalizing again.
540       StringRef OrigName =
541           ModuleSummaryIndex::getOriginalNameBeforePromote(GV.getName());
542       std::string OrigId = GlobalValue::getGlobalIdentifier(
543           OrigName, GlobalValue::InternalLinkage,
544           TheModule.getSourceFileName());
545       const auto &GS = DefinedGlobals.find(GlobalValue::getGUID(OrigId));
546       if (GS == DefinedGlobals.end()) {
547         // Also check the original non-promoted non-globalized name. In some
548         // cases a preempted weak value is linked in as a local copy because
549         // it is referenced by an alias (IRLinker::linkGlobalValueProto).
550         // In that case, since it was originally not a local value, it was
551         // recorded in the index using the original name.
552         // FIXME: This may not be needed once PR27866 is fixed.
553         const auto &GS = DefinedGlobals.find(GlobalValue::getGUID(OrigName));
554         assert(GS != DefinedGlobals.end());
555         Linkage = GS->second->linkage();
556       } else {
557         Linkage = GS->second->linkage();
558       }
559     } else
560       Linkage = GS->second->linkage();
561     return !GlobalValue::isLocalLinkage(Linkage);
562   };
563 
564   // FIXME: See if we can just internalize directly here via linkage changes
565   // based on the index, rather than invoking internalizeModule.
566   llvm::internalizeModule(TheModule, MustPreserveGV);
567 }
568 
569 // Automatically import functions in Module \p DestModule based on the summaries
570 // index.
571 //
572 bool FunctionImporter::importFunctions(
573     Module &DestModule, const FunctionImporter::ImportMapTy &ImportList,
574     bool ForceImportReferencedDiscardableSymbols) {
575   DEBUG(dbgs() << "Starting import for Module "
576                << DestModule.getModuleIdentifier() << "\n");
577   unsigned ImportedCount = 0;
578 
579   // Linker that will be used for importing function
580   Linker TheLinker(DestModule);
581   // Do the actual import of functions now, one Module at a time
582   std::set<StringRef> ModuleNameOrderedList;
583   for (auto &FunctionsToImportPerModule : ImportList) {
584     ModuleNameOrderedList.insert(FunctionsToImportPerModule.first());
585   }
586   for (auto &Name : ModuleNameOrderedList) {
587     // Get the module for the import
588     const auto &FunctionsToImportPerModule = ImportList.find(Name);
589     assert(FunctionsToImportPerModule != ImportList.end());
590     std::unique_ptr<Module> SrcModule = ModuleLoader(Name);
591     assert(&DestModule.getContext() == &SrcModule->getContext() &&
592            "Context mismatch");
593 
594     // If modules were created with lazy metadata loading, materialize it
595     // now, before linking it (otherwise this will be a noop).
596     SrcModule->materializeMetadata();
597     UpgradeDebugInfo(*SrcModule);
598 
599     auto &ImportGUIDs = FunctionsToImportPerModule->second;
600     // Find the globals to import
601     DenseSet<const GlobalValue *> GlobalsToImport;
602     for (Function &F : *SrcModule) {
603       if (!F.hasName())
604         continue;
605       auto GUID = F.getGUID();
606       auto Import = ImportGUIDs.count(GUID);
607       DEBUG(dbgs() << (Import ? "Is" : "Not") << " importing function " << GUID
608                    << " " << F.getName() << " from "
609                    << SrcModule->getSourceFileName() << "\n");
610       if (Import) {
611         F.materialize();
612         if (EnableImportMetadata) {
613           // Add 'thinlto_src_module' metadata for statistics and debugging.
614           F.setMetadata(
615               "thinlto_src_module",
616               llvm::MDNode::get(
617                   DestModule.getContext(),
618                   {llvm::MDString::get(DestModule.getContext(),
619                                        SrcModule->getSourceFileName())}));
620         }
621         GlobalsToImport.insert(&F);
622       }
623     }
624     for (GlobalVariable &GV : SrcModule->globals()) {
625       if (!GV.hasName())
626         continue;
627       auto GUID = GV.getGUID();
628       auto Import = ImportGUIDs.count(GUID);
629       DEBUG(dbgs() << (Import ? "Is" : "Not") << " importing global " << GUID
630                    << " " << GV.getName() << " from "
631                    << SrcModule->getSourceFileName() << "\n");
632       if (Import) {
633         GV.materialize();
634         GlobalsToImport.insert(&GV);
635       }
636     }
637     for (GlobalAlias &GA : SrcModule->aliases()) {
638       if (!GA.hasName())
639         continue;
640       auto GUID = GA.getGUID();
641       auto Import = ImportGUIDs.count(GUID);
642       DEBUG(dbgs() << (Import ? "Is" : "Not") << " importing alias " << GUID
643                    << " " << GA.getName() << " from "
644                    << SrcModule->getSourceFileName() << "\n");
645       if (Import) {
646         // Alias can't point to "available_externally". However when we import
647         // linkOnceODR the linkage does not change. So we import the alias
648         // and aliasee only in this case. This has been handled by
649         // computeImportForFunction()
650         GlobalObject *GO = GA.getBaseObject();
651         assert(GO->hasLinkOnceODRLinkage() &&
652                "Unexpected alias to a non-linkonceODR in import list");
653 #ifndef NDEBUG
654         if (!GlobalsToImport.count(GO))
655           DEBUG(dbgs() << " alias triggers importing aliasee " << GO->getGUID()
656                        << " " << GO->getName() << " from "
657                        << SrcModule->getSourceFileName() << "\n");
658 #endif
659         GO->materialize();
660         GlobalsToImport.insert(GO);
661         GA.materialize();
662         GlobalsToImport.insert(&GA);
663       }
664     }
665 
666     // Link in the specified functions.
667     if (renameModuleForThinLTO(*SrcModule, Index, &GlobalsToImport))
668       return true;
669 
670     if (PrintImports) {
671       for (const auto *GV : GlobalsToImport)
672         dbgs() << DestModule.getSourceFileName() << ": Import " << GV->getName()
673                << " from " << SrcModule->getSourceFileName() << "\n";
674     }
675 
676     // Instruct the linker that the client will take care of linkonce resolution
677     unsigned Flags = Linker::Flags::None;
678     if (!ForceImportReferencedDiscardableSymbols)
679       Flags |= Linker::Flags::DontForceLinkLinkonceODR;
680 
681     if (TheLinker.linkInModule(std::move(SrcModule), Flags, &GlobalsToImport))
682       report_fatal_error("Function Import: link error");
683 
684     ImportedCount += GlobalsToImport.size();
685   }
686 
687   NumImported += ImportedCount;
688 
689   DEBUG(dbgs() << "Imported " << ImportedCount << " functions for Module "
690                << DestModule.getModuleIdentifier() << "\n");
691   return ImportedCount;
692 }
693 
694 /// Summary file to use for function importing when using -function-import from
695 /// the command line.
696 static cl::opt<std::string>
697     SummaryFile("summary-file",
698                 cl::desc("The summary file to use for function importing."));
699 
700 static void diagnosticHandler(const DiagnosticInfo &DI) {
701   raw_ostream &OS = errs();
702   DiagnosticPrinterRawOStream DP(OS);
703   DI.print(DP);
704   OS << '\n';
705 }
706 
707 /// Parse the summary index out of an IR file and return the summary
708 /// index object if found, or nullptr if not.
709 static std::unique_ptr<ModuleSummaryIndex> getModuleSummaryIndexForFile(
710     StringRef Path, std::string &Error,
711     const DiagnosticHandlerFunction &DiagnosticHandler) {
712   std::unique_ptr<MemoryBuffer> Buffer;
713   ErrorOr<std::unique_ptr<MemoryBuffer>> BufferOrErr =
714       MemoryBuffer::getFile(Path);
715   if (std::error_code EC = BufferOrErr.getError()) {
716     Error = EC.message();
717     return nullptr;
718   }
719   Buffer = std::move(BufferOrErr.get());
720   ErrorOr<std::unique_ptr<object::ModuleSummaryIndexObjectFile>> ObjOrErr =
721       object::ModuleSummaryIndexObjectFile::create(Buffer->getMemBufferRef(),
722                                                    DiagnosticHandler);
723   if (std::error_code EC = ObjOrErr.getError()) {
724     Error = EC.message();
725     return nullptr;
726   }
727   return (*ObjOrErr)->takeIndex();
728 }
729 
730 static bool doImportingForModule(Module &M, const ModuleSummaryIndex *Index) {
731   if (SummaryFile.empty() && !Index)
732     report_fatal_error("error: -function-import requires -summary-file or "
733                        "file from frontend\n");
734   std::unique_ptr<ModuleSummaryIndex> IndexPtr;
735   if (!SummaryFile.empty()) {
736     if (Index)
737       report_fatal_error("error: -summary-file and index from frontend\n");
738     std::string Error;
739     IndexPtr =
740         getModuleSummaryIndexForFile(SummaryFile, Error, diagnosticHandler);
741     if (!IndexPtr) {
742       errs() << "Error loading file '" << SummaryFile << "': " << Error << "\n";
743       return false;
744     }
745     Index = IndexPtr.get();
746   }
747 
748   // First step is collecting the import list.
749   FunctionImporter::ImportMapTy ImportList;
750   ComputeCrossModuleImportForModule(M.getModuleIdentifier(), *Index,
751                                     ImportList);
752 
753   // Next we need to promote to global scope and rename any local values that
754   // are potentially exported to other modules.
755   if (renameModuleForThinLTO(M, *Index, nullptr)) {
756     errs() << "Error renaming module\n";
757     return false;
758   }
759 
760   // Perform the import now.
761   auto ModuleLoader = [&M](StringRef Identifier) {
762     return loadFile(Identifier, M.getContext());
763   };
764   FunctionImporter Importer(*Index, ModuleLoader);
765   return Importer.importFunctions(M, ImportList,
766                                   !DontForceImportReferencedDiscardableSymbols);
767 }
768 
769 namespace {
770 /// Pass that performs cross-module function import provided a summary file.
771 class FunctionImportLegacyPass : public ModulePass {
772   /// Optional module summary index to use for importing, otherwise
773   /// the summary-file option must be specified.
774   const ModuleSummaryIndex *Index;
775 
776 public:
777   /// Pass identification, replacement for typeid
778   static char ID;
779 
780   /// Specify pass name for debug output
781   const char *getPassName() const override { return "Function Importing"; }
782 
783   explicit FunctionImportLegacyPass(const ModuleSummaryIndex *Index = nullptr)
784       : ModulePass(ID), Index(Index) {}
785 
786   bool runOnModule(Module &M) override {
787     if (skipModule(M))
788       return false;
789 
790     return doImportingForModule(M, Index);
791   }
792 };
793 } // anonymous namespace
794 
795 PreservedAnalyses FunctionImportPass::run(Module &M,
796                                           ModuleAnalysisManager &AM) {
797   if (!doImportingForModule(M, Index))
798     return PreservedAnalyses::all();
799 
800   return PreservedAnalyses::none();
801 }
802 
803 char FunctionImportLegacyPass::ID = 0;
804 INITIALIZE_PASS(FunctionImportLegacyPass, "function-import",
805                 "Summary Based Function Import", false, false)
806 
807 namespace llvm {
808 Pass *createFunctionImportPass(const ModuleSummaryIndex *Index = nullptr) {
809   return new FunctionImportLegacyPass(Index);
810 }
811 }
812